DeFLOCNet: Deep Image Editing via Flexible Low-Level Controls

Input (b) Ours (c) Input (d) Ours (e) Input (f) Ours (g) Original Figure 1. Diversified image editing results under free-form input configurations. There is a face or a natural image with arbitrary hole regions. Without any user inputs shown in (a), our DeFLOCNet automatically fills hole regions shown in (b), which is the same as the image inpainting scenario. Given additional low-level controls (e.g., coarse sketch lines in (c), both lines and colors in (e)), our DeFLOCNet injects these controls directly into structure generation blocks for both user-intended and visually pleasant content creations in (d) and (f).